SkylakeX teasers (aka prime95 29.5) - Page 14

kriesel · 2019-01-09, 21:26

Quote:

Originally Posted by Prime95

Yes, that was a long shot. Thanks for trying.

Did you ever check if 29.5b6 hangs in day-to-day normal use (1 or 3 workers with or without hyperthreading). I think it should hang there too.

Updates were needed done anyway, no problem.
Test update: v29.4b8 completed a benchmark 1024k-32768 HT and not, on 1-3,6 workers, first try, no stall on the same i7-8750H hardware and Windows 10 install. That was without Mfakto on the igp or Mfaktc on the gtx1050Ti in it.

My recollection is that both V29.4b8 and v29.5b6 were stable in ordinary use in Win 10; there's no assignment or work in progress in the v29.5b5 folder. One worker, 6 cores, no HT during primality tests. I've postponed work on the V29.4b8 first-PRP test in the 82M range, to run a day to completion of a 51M LL DC in V29.5b6, and will check it for any stall or other quirk along the way. Mfakto and Mfaktc have been resumed and so far after just 30 minutes it all seems stable in normal use.

Memory limit is 8192MB in all of them The total installed memory is 16384 MB on this system.

kriesel · 2019-01-09, 23:22

George,

I went back through the thread and put together a summary chart from data I had myself or could find there. Maybe you've already done this and a lot more. Anyway, I note the time element; 29.4b8 fine, 29.5b3, 5, 6, stalling on some systems, both Windows and linux apparently, and on varying hardware.
Behavior of b1, b2, unknown; b4 presumed to stall also.

Maybe the time element might provide a clue? Wouldn't it be interesting if b1 was ok regarding the benchmark stall? I'm willing to put some early versions on my "reliable" system for a quick test for benchmark stall, if they're made available/accessible. For b1 or b2, I would need a download link in PM or posted here, or in email a link or zip file for each.

Folks who have hit the stall issue and who'd like to share a few more of the tabulated parameters, please do.

We'll help you corner and cage this gremlin however we can.

kladner · 2019-01-10, 17:56

I tried disabling HT. I then ran repeated (haven't counted) benchmarks for 2560K. I did not encounter any stalls. The only thing I noticed is that with HT runs slightly faster.
Typical peak with HT:
FFTlen=2560K, Type=3, Arch=4, Pass1=1024, Pass2=2560, clm=1 (4 cores, 1 worker): 2.45 ms. Throughput: 408.35 iter/sec.

Without HT:
FFTlen=2560K, Type=3, Arch=4, Pass1=640, Pass2=4096, clm=1 (4 cores, 1 worker): 2.51 ms. Throughput: 398.70 iter/sec.
However, the latter is an outlier which only occurred in one run. Otherwise, it did not make it out of the high 380's per second.

Prime95 · 2019-01-10, 23:05

Finally! I got a crash while benchmarking under linux on Skylake-X.

I found a small window of vulnerability when calling gwdone (the routine that says I'm done FFTing - close threads and free memory). This routine is called far more often during a benchmark as opposed to daily work.

Could this vulnerability also cause the hangs? Maybe. I didn't study all the ramifications of the vulnerability.

I coded up a fix. Can the Windows users try again with this executable, thanks:
https://www.dropbox.com/s/sc4ib5v4f4...ime95.zip?dl=0

kriesel · 2019-01-11, 03:48

Quote:

Originally Posted by Prime95

Finally! I got a crash while benchmarking under linux on Skylake-X.

I found a small window of vulnerability when calling gwdone (the routine that says I'm done FFTing - close threads and free memory). This routine is called far more often during a benchmark as opposed to daily work.

Could this vulnerability also cause the hangs? Maybe. I didn't study all the ramifications of the vulnerability.

I coded up a fix. Can the Windows users try again with this executable, thanks:
https://www.dropbox.com/s/sc4ib5v4f4...ime95.zip?dl=0

By George, I believe you've got it!

My test i7-8750H completed a DC in prime95 29.5b6, https://www.mersenne.org/report_expo...1499457&full=1, then confirmed benchmark hangs in build 3 and 4 also, and ran an extensive v29.5 build 7 benchmark to completion (1-3,6 workers, HT and no, 1024k-32768k)
Time for a linux build 7 to testers?

kladner · 2019-01-11, 04:17

Thanks, George!
I will put it to work at once.

EDIT:
It ran fine overnight, with an Auto Bench included. I should note that the bench was for 2688K. When I have the time today I will run some 2560K tests.

Falkentyne · 2019-01-11, 19:13

Quote:

Originally Posted by Prime95

Finally! I got a crash while benchmarking under linux on Skylake-X.

I found a small window of vulnerability when calling gwdone (the routine that says I'm done FFTing - close threads and free memory). This routine is called far more often during a benchmark as opposed to daily work.

Could this vulnerability also cause the hangs? Maybe. I didn't study all the ramifications of the vulnerability.

I coded up a fix. Can the Windows users try again with this executable, thanks:
https://www.dropbox.com/s/sc4ib5v4f4...ime95.zip?dl=0

Found a bug (in both build 6 and 7, I did not test previous ones).
The 14K fft size fails instantly on all cores with a roundoff error (.5, expected less than .4)

No problem on the last stable release 29.4 build 8.

Prime95 · 2019-01-11, 19:59

Quote:

Originally Posted by Falkentyne

Found a bug (in both build 6 and 7, I did not test previous ones).
The 14K fft size fails instantly on all cores with a roundoff error (.5, expected less than .4)

No problem on the last stable release 29.4 build 8.

Correct, debugging now. Note, this only affects SSE2 FFTs. There is no AVX, FMA3, or AVX-512 FFT of length 14K.

Prime95 · 2019-01-12, 03:35

Quote:

Originally Posted by Falkentyne

Found a bug (in both build 6 and 7, I did not test previous ones).
The 14K fft size fails instantly on all cores with a roundoff error (.5, expected less than .4)

The new torture test feature to run a weaker test had a bug. It was selecting exponents that were too big for the FFT size. A new Windows executable for you to test:

https://www.dropbox.com/s/sc4ib5v4f4...ime95.zip?dl=0

tshinozk · 2019-01-12, 06:36

>I coded up a fix. Can the Windows users try again with this executable, thanks:

The hang is fixed in my machine.
I finished the benchmarks with no problems.
Thanks.

pepi37 · 2019-01-12, 16:09

Where I can download latest version for linux64 bit?
And what is latest version beta6 or beta7?

2019-01-11, 04:17	#149
kladner "Kieren" Jul 2011 In My Own Galaxy! 2·3·1,693 Posts	Thanks, George! I will put it to work at once. EDIT: It ran fine overnight, with an Auto Bench included. I should note that the bench was for 2688K. When I have the time today I will run some 2560K tests. Last fiddled with by kladner on 2019-01-11 at 12:42

2019-01-10, 17:56	#146
kladner "Kieren" Jul 2011 In My Own Galaxy! 23656₈ Posts	I tried disabling HT. I then ran repeated (haven't counted) benchmarks for 2560K. I did not encounter any stalls. The only thing I noticed is that with HT runs slightly faster. Typical peak with HT: FFTlen=2560K, Type=3, Arch=4, Pass1=1024, Pass2=2560, clm=1 (4 cores, 1 worker): 2.45 ms. Throughput: 408.35 iter/sec. Without HT: FFTlen=2560K, Type=3, Arch=4, Pass1=640, Pass2=4096, clm=1 (4 cores, 1 worker): 2.51 ms. Throughput: 398.70 iter/sec. However, the latter is an outlier which only occurred in one run. Otherwise, it did not make it out of the high 380's per second.

2019-01-10, 23:05	#147
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 7,411 Posts	Finally! I got a crash while benchmarking under linux on Skylake-X. I found a small window of vulnerability when calling gwdone (the routine that says I'm done FFTing - close threads and free memory). This routine is called far more often during a benchmark as opposed to daily work. Could this vulnerability also cause the hangs? Maybe. I didn't study all the ramifications of the vulnerability. I coded up a fix. Can the Windows users try again with this executable, thanks: https://www.dropbox.com/s/sc4ib5v4f4...ime95.zip?dl=0

2019-01-12, 06:36	#153
tshinozk Nov 2012 23 Posts	>I coded up a fix. Can the Windows users try again with this executable, thanks: The hang is fixed in my machine. I finished the benchmarks with no problems. Thanks.

2019-01-12, 16:09	#154
pepi37 Dec 2011 After milion nines:) 3·11·43 Posts	Where I can download latest version for linux64 bit? And what is latest version beta6 or beta7?